ETRI-Knowledge Sharing Plaform

ENGLISH

성과물

논문 검색
구분 SCI
연도 ~ 키워드

상세정보

학술대회 An Automatic Database Generation Algorithm for Local Optimization of CNN Object Detector for Edge Devices
Cited 0 time in scopus Download 3 time Share share facebook twitter linkedin kakaostory
저자
이준구, 백장운
발행일
202011
출처
International Conference on Consumer Electronics (ICCE) 2020 : Asia, pp.364-366
DOI
https://dx.doi.org/10.1109/ICCE-Asia49877.2020.9276780
협약과제
20ZD1100, 대경권 지역산업 기반 ICT 융합기술 고도화 지원사업, 문기영
초록
Recently, the demand for the use of deep learning algorithm in edge devices is increasing. Deep learning algorithm needs high computation power and large memory resources. However, edge devices require high accuracy and real-time performance with limited resources. In order to overcome this problem, the lightweight shallow networks have been proposed, but their accuracy is much lower than the existing dense networks. We observed that edge devices such as surveillance and security CCTVs are located at the fixed area. We focused these environments where local optimization, which retrains the object detector using a new local database, is very effective to improve the detection accuracy. Local optimization needs additional annotation work for local training database, which is tiresome and time-consuming. We proposed an automatic database generation algorithm for local optimization, which uses a pre-trained object detector and a background model. The proposed algorithm generates the training images by overlaying the extracted objects from the object detector on the background image from background modelling.
키워드
automatic annotation, edge computing, Local optimization, object detection
KSP 제안 키워드
Automatic Annotation, Background image, Background modelling, Computation power, Database generation, Dense network, Detection accuracy, Edge devices, Generation algorithm, High accuracy, Large memory